Exchange rate forecasting using economic models and technical trading rules

Abstract

The use of technical analysis by practitioners in the foreign exchange market contrasts with the ongoing debate among academics on the poor predictive ability of macroeconomic variables. This paper compares these two methods by constructing pools of economic models and technical trading rules and evaluates their in-sample and out-of-sample performance both locally and globally. Results suggest the presence of local forecastability that is overlooked when relying on global measures of predictability. The local predictability is captured using a rolling model selection approach to generate aggregate forecasts across separate pools of economic models and technical trading rules as well as both combined. The out-of-sample results for our aggregate forecasts using pools of economic models fail to beat the random walk as do pools of technical trading models. However combining the two pools of models results in forecasts that beat the random walk for four out of the six sample currencies. This result suggests that exchange rate forecasts can be improved by pooling both sets of models.

Keywords:

• FX excess returns
• out-of-sample predictability
• time series
• econometric models

JEL Classifications:

• F31
• F37
• G11
• G17

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 One can think of globally as referring to the entire out-of-sample period and locally as referring to part of the out-of-sample period. See Giacomini and Rossi's (2010) motivating example for more details.

2 For a recent study that finds TTRs profitable see Alanazi (2020).

3 The money supply is denoted by m which includes physical money, demand deposits, checking accounts and negotiable order of withdrawal accounts.

4 Following Rossi (2013) we use a one-sided moving average seasonal adjustment filter instead of a two-sided one, as forecasts generated based on two-sided filtered data use information which would have not been available at the time of the forecast. We also take the log of all the variables, except overnight interest rates.

5 We follow Belsley, Kuh, and Welsch (1980) and use the variance inflation factor (VIF) to check for multi-collinearity between economic variables. We find moderate multi-collinearity consistent with the Gauss–Markov assumptions.

6 This implies that our pool of models includes UIRP, PPP, Taylor rule fundamentals, the monetary model, and Balassa–Samuelson model.

7 See Appendix 1 for sample plots for an arbitrary economic model and TTR.

8 See Appendix 2 for further details on the bootstrap procedure. Of course, it is also possible to consider averaging across other dimensions, for example estimation procedures where there is specification uncertainty. See Steel (2020).

9 Evidence suggests that the 1/n scheme seems to perform better than estimated optimal forecast combinations, particularly when the number of models is large. See Smith and Wallis (2009) and Claeskens et al. (2016).

10 Appendix 3 shows the proportion of models selected for the economic and TTR pools separately.

11 The inability to beat the random walk benchmark for USD/GBP and USD/JPY is not surprising in light of the absence of local predictability from the fluctuation test. For robustness we confirm the results of Table 3 using the test of Clark and West (2007). Further we forecasted using Bayesian methods (dynamic model averaging and dynamic model selection, see Raftery, Karny, and Ettler 2010) and calculated the cumulative sum of squared out-of-sample forecast errors (versus the random walk) as per (Welch and Goyal 2007). Results suggest superior performance against the random walk for the same currency pairs that exhibited significance in Table 3. Results not reported to conserve space.

12 To reflect the higher transaction costs in the earlier part of our sample, we follow Neely, Weller, and Ulrich (2009) by varying transaction costs for switching from a long to a short position from 10 basis points in the 1970s decreasing to just 2 basis points in the latter part of our sample. For completeness we also examined the Sharpe ratio of the buy-and-hold strategy for each currency pair (unreported). In all cases the Sharpe ratio was greater than those from the composite forecasts.

Additional information

Notes on contributors

Nima Zarrabi

Nima Zarrabi is a UK Equities Investment Analyst at Invesco. Previous to this, he was an Investment Risk Manager conducting quantitative analysis across different asset classes and strategies. His research focuses on international finance and applied econometrics. His research has been published in the International Review of Financial Analysis.

Stuart Snaith

Stuart Snaith is an Associate Professor at the Peter B. Gustavson School of Business. His research focuses primarily on international finance. He has published articles in peer-reviewed journals such as the Journal of Banking and Finance, Journal of Futures Markets, International Review of Financial Analysis, and Economics Letters.

Jerry Coakley

Jerry Coakley is Professor of Finance at Essex Business School. His research focuses on behavioural corporate finance, FinTech, crowdfunding and entrepreneurship. He has published articles in peer-reviewed journals such as the Economic Journal, Journal of Corporate Finance, Journal of Banking & Finance, amongst others. He is an Associate Editor of the European Journal of Finance.